Detecção de defeitos em caldeiras de recuperação química

• Ano de defesa: 2008
• Autor(a) principal: Martinelli, Sérgio Henrique Saavedra
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual de Maringá; Brasil; Programa de Pós-Graduação em Engenharia Química; UEM; Maringá, PR; Departamento de Engenharia Química
• Palavras-chave em Português: Caldeiras; Detecção de defeitos; Processos industriais; Controle; Caldeiras de recuperação química; Brasil.; Engenharias; Engenharia Química
• Link de acesso: http://repositorio.uem.br:8080/jspui/handle/1/3633

The industrial processes are becoming even more unmanned and dependent on the control components. The way to improve the processes reliability is to guarantee the reliability and robustness of these control components, however a defect-free process can not be guaranteed. Defects in control systems can show up in an abrupt or incipient way. The last one is extremely difficult to detect as its effects are covered up by the handled variables of the control loops. In this thesis two artificial neural networks are used to detect incipient defects in a chemical recovery boiler in the kraft pulp and paper manufacturing. A multilayer perceptron neural network is used to model the process behavior. The neural networks outputs are compared to the measurements in the process generating residues. Other multilayer perceptron neural network is used to classify the residues. The training of this classifying neural network was carried out with faulty data generated by inserting errors in the data of process variables. The method was applied in the control loops: primary, secondary and tertiary air flow, water temperature of smelt spout cooling and the furnace pressure. After the training stage, the two neural networks were used for the continuous process monitoring. The method presented false alarms rate of 0.2 to 5.0%, to the control loops analyzed. The rate of defects detection was of 89 to 97% to errors of 9 to 16% inserted at the process variables. The results presented demonstrated the method can detect correctly defects in non-trained pattern sets and constitute an alternative and reliable tool to operational support.